Conference Paper Statistical Methods for Attrition and Non-Response in Social Surveys Conference
The Long-Term Effectiveness of Procedures for Minimising Attrition on Longitudinal Surveys
28 May 2004
A commonly accepted view is that longitudinal surveys are both expensive to run and present data quality problems if levels of attrition from the survey are high, especially where there is evidence of differential patterns of attrition for particular groups within the sample which may introduce some bias into estimates from the data. There are also concerns about maintaining overall sample sizes for those with continuous longitudinal records as well as the size of sub-groups of interest within the longitudinal sample. Longitudinal surveys therefore go to great lengths to try and keep respondents participating in the survey using a variety of means to do so. During fieldwork, these typically involve a range of tracing procedures, refusal conversion programmes or the re-issuing of non-contacts for further attempts. All of these procedures are costly to implement even though they affect a relatively small proportion of the sample at any given survey wave. Nonetheless they are considered central elements for ensuring the long term viability of a longitudinal survey. Despite a large literature on attrition there is little evidence on the actual value of these types of procedures for the quality of longitudinal data and it is this question which primarily concerns us in this paper.
The paper uses data from the British Household Panel Survey (BHPS) from 1994 to 2002 to assess the long term effectiveness of procedures for minimising attrition in maintaining sample sizes and for data quality. We focus here on two procedures: refusal conversion, and attempts to 'track' mover households to a new address. Since 1994, the BHPS has collected information on these processes and these data allow us to examine how successful the procedures have been in maintaining the viability of the sample over time.
We first examine the longevity of successful refusal conversion. Are we simply postponing drop out for a wave or two, or is the procedure effective at retaining units in the sample for many waves? We then compare sample sizes and sample characteristics, both overall and for subgroups, across multiple waves, with and without refusal conversion. With regard to tracking movers to new addresses, we examine success rates and their impacts on sample sizes and sample characteristics. We also examine the combined effects of both procedures. Finally, we make a preliminary assessment of the effects of the procedures on data quality through comparing key survey estimates.